Fully automating LI-RADS on MRI with deep learning-guided lesion segmentation, feature characterization, and score...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning holds significant potential for radiology applications.
• The Liver Imaging Reporting and Data System (LI-RADS) is crucial for liver lesion characterization.
• Integrating AI into LI-RADS can enhance diagnostic accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate a fully automatic deep learning solution for the LI-RADS system.
• To investigate the model's performance in liver lesion segmentation and classification.
• To introduce a feature characterization step to improve diagnostic performance and model explainability.

Main Methods:

• A UNet++-based segmentation approach was employed using 33,078 images from 111 patients.
• A novel feature characterization network with multi-task learning and joint training was introduced before classification.
• An inference module was used to generate the final LI-RADS score.

Main Results:

• The liver lesion segmentation achieved a median DICE score of 0.879 with high precision (>0.9) and recall (0.7-0.9).
• Patient-level segmentation performance showed excellent results with (precision, recall) approximately (1, 0.9).
• Lesion classification achieved an overall accuracy of 76%, with most misclassifications occurring in neighboring LI-RADS categories.

Conclusions:

• The proposed deep learning framework with feature characterization significantly improves diagnostic performance in LI-RADS.
• The feature characterization step enhances model explainability by providing intermediate results.
• This AI-driven approach offers a promising tool for automated liver lesion assessment.